Fast Parameter Learning for Markov Logic Networks Using Bayes Nets

Abstract

Markov Logic Networks (MLNs) are a prominent statistical relational model that have been proposed as a unifying framework for statistical relational learning. As part of this unification, their authors proposed methods for converting other statistical relational learners into MLNs. For converting a first order Bayes net into an MLN, it was suggested to moralize the Bayes net to obtain the structure of the MLN and then use the log of the conditional probability table entries to calculate the weight of the clauses. This conversion is exact for converting propositional Markov networks to propositional Bayes nets however, it fails to perform well for the relational case. We theoretically analyze this conversion and introduce new methods of converting a Bayes net into an MLN. An extended imperial evaluation on five datasets indicates that our conversion method outperforms previous methods.